vLLM has a Weakness in MultiModalHasher Image Hashing Implementation

Description

Summary

In the file vllm/multimodal/hasher.py, the MultiModalHasher class has a security and data integrity issue in its image hashing method. Currently, it serializes PIL.Image.Image objects using only obj.tobytes(), which returns only the raw pixel data, without including metadata such as the image’s shape (width, height, mode). As a result, two images of different sizes (e.g., 30x100 and 100x30) with the same pixel byte sequence could generate the same hash value. This may lead to hash collisions, incorrect cache hits, and even data leakage or security risks.

Details

  • Affected file: vllm/multimodal/hasher.py
  • Affected method: MultiModalHasher.serialize_item
    https://github.com/vllm-project/vllm/blob/9420a1fc30af1a632bbc2c66eb8668f3af41f026/vllm/multimodal/hasher.py#L34-L35
  • Current behavior: For Image.Image instances, only obj.tobytes() is used for hashing.
  • Problem description: obj.tobytes() does not include the image’s width, height, or mode metadata.
  • Impact: Two images with the same pixel byte sequence but different sizes could be regarded as the same image by the cache and hashing system, which may result in:
    • Incorrect cache hits, leading to abnormal responses
    • Deliberate construction of images with different meanings but the same hash value

Recommendation

In the serialize_item method, serialization of Image.Image objects should include not only pixel data, but also all critical metadata—such as dimensions (size), color mode (mode), format, and especially the info dictionary. The info dictionary is particularly important in palette-based images (e.g., mode 'P'), where the palette itself is stored in info. Ignoring info can result in hash collisions between visually distinct images with the same pixel bytes but different palettes or metadata. This can lead to incorrect cache hits or even data leakage.

Summary:
Serializing only the raw pixel data is insecure. Always include all image metadata (size, mode, format, info) in the hash calculation to prevent collisions, especially in cases like palette-based images.

Impact for other modalities
For the influence of other modalities, since the video modality is transformed into a multi-dimensional array containing the length, width, time, etc. of the video, the same problem exists due to the incorrect sequence of numpy as well.

For audio, since the momo function is not enabled in librosa.load, the loaded audio is automatically encoded into single channels by librosa and returns a one-dimensional array of numpy, thus keeping the structure of numpy fixed and not affected by this issue.

Fixes

  • https://github.com/vllm-project/vllm/pull/17378

Basic information

Type
reviewed
Severity
medium
Advisory on GitHub
Open advisory ↗
Repository advisory
Open repository advisory ↗
Source code
Browse source ↗
Published (advisory)
2025-05-28 18:03:41 UTC
Updated
2025-05-29 21:36:28 UTC
GitHub reviewed
2025-05-28 18:03:41 UTC
NVD published
2025-05-29

EPSS Score

Score Percentile
0.23% 45.76%

CVSS Scores

Base score Version Severity Vector
4.2 3.1
CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:L/I:N/A:L Click to expand
Attack vector (AV:N)
Could be attacked over the internet or any normal routed network—not just someone sitting at the machine.
Attack complexity (AC:H)
Even with access, the exploit needs extra luck, timing, or a fussy environment to actually work.
Privileges required (PR:L)
A normal user session is enough; they don’t have to be admin.
User interaction (UI:N)
Nobody has to click “OK” or open a trap file; it can work without a victim helping.
Scope (S:U)
Damage stays in the same “trust bubble” as the broken component—no big spill into unrelated systems.
Confidentiality (C:L)
Some sensitive info could get out, but not a total data dump.
Integrity (I:N)
Data isn’t meaningfully altered or forged.
Availability (A:L)
Might cause slowdowns, glitches, or partial disruption—not a full brick.

Identifiers

CWEs

CWE id Name
CWE-1023 Incomplete Comparison with Missing Factors
CWE-1288 Improper Validation of Consistency within Input

Credits

  • kexinoh (reporter)
  • DarkLight1337 (remediation_developer)
  • russellb (coordinator)

Affected packages (1)

Vulnerable version ranges and first patched releases as published by GitHub.

Ecosystem Package Vulnerable range First patched Vulnerable functions
pip vllm >= 0.7.0, < 0.9.0 0.9.0

References

cvelogic Threat Intelligence